Contact

Ymir Labs · AI research lab

More useful intelligence from every unit of compute.

We research how to get far more out of the compute that already exists. We’re building Candid: AI agents designed to get cheaper on work they’ve done before, take corrections partway through a long task, and run as small models on ordinary hardware, even offline.

Our researchSee REN’s results

Latest measured result

Measured

REN, round five, against its base model

  • Tool-using agent turns answered correctly0.740 → 1.000

    Clearly better

  • Handles the interruption0.786 → 0.833

    Within noise: not a win yet

REN is a solo proof of concept, not Candid. One win, one not yet, both published. All eight results

Stage
Early-stage. Pre-product, pre-revenue.
Team
One person, full-time.
Facts checked

01Thesis

Frontier AI has grown by spending more.

Most of the progress in frontier AI has come from spending more compute. It works. It also makes AI expensive, slow and concentrated.

  • Expensive

    Most AI pays full price for every task, even one it has solved a thousand times.

  • Slow

    More compute per answer means more waiting per answer, and more waiting between ideas.

  • Concentrated

    Capable AI gathers where the biggest compute budgets are, rather than where people need it.

There is a great deal of room to get more out of the compute that already exists. That room is what we study.

02The metric

The number we care about.

Useful work per unit of compute. We intend to report it for what we build, alongside a second number that keeps us honest.

  1. 1

    Useful work per unit of compute

    How much real work gets done for the compute it costs. The thing we’re trying to raise.

    Planned
  2. 2

    Predictions that hold up

    The share of our predictions that hold up when checked for real. The thing that tells you whether to believe the first.

    Planned

Neither number exists for Candid yet, because Candid doesn’t exist yet. When they do, they’ll be published here, whichever way they point.

03What we’re building

Candid: AI agents that get cheaper as they work.

Candid is the lab’s flagship product. It is in research and design, and will be built once the lab is funded. This is what it is designed to do.

  1. Gets cheaper with experience

    In research

    Most AI pays full price for every task, even one it has solved a thousand times. Candid is designed to get faster and cheaper on the kinds of work it has already done.

  2. Correctable mid-task

    In research

    Change your mind halfway through a large job, say a big software build, and Candid updates what needs to change, checks what else that affects, and carries on. No starting over. No quietly breaking earlier work.

  3. Runs where people are

    In research

    Small versions designed to run on ordinary laptops, offline, so data can stay on the device and costs stay predictable. Larger versions in the cloud or on a company’s own servers.

  4. Tested, not just described

    Planned

    A virtual lab for electronics, robotics and engineering experiments, with key predictions checked against real measurements on a physical bench, and every result published with what’s needed to reproduce it. The bench doesn’t exist yet.

  5. Open where it matters

    Planned

    We plan to release small models openly.

More on Candid

04Evidence

What one person can do on free GPUs.

REN is a solo proof of concept by the founder: a model trained, evaluated and published rigorously with zero budget. It is not Candid, and Candid is not built on it. It shows how we measure.

  • Base Qwen3.5-9B
  • REN, round five
  • Scale 0 to 1
Four of eight results from REN’s fifth round, on a held-out test set against the untouched base model. The last row is the one REN was trained for, and it is not a win yet. Measured
MeasureBaseRENVerdict
Terminal tasks completedEnd state checked · n=420.85736/421.00042/42Clearly better
Quick judgementsAccuracy · n=3000.6770.830Clearly better
CalibrationBrier score, lower is better · n=3000.4190.252Clearly better
Handles the interruptionn=420.78633/420.83335/42Within noise: not a win yet

We print the row that didn’t clear the bar with the same care as the rows that did. A lab that hides its noise can’t be trusted with its signal.

All eight resultsPublic training log

05Stage

Where we are, plainly.

Stage
Early-stage. Pre-product and pre-revenue.
Candid
In research and design. It will be built once the lab is funded.
Team
One person, full-time: the founder, Oviezibe Gift Jack. About the lab
Customers and users
None yet.
Partners and funding
None to date.
Company
Ymir Labs is the research lab of Nutin Technologies, which will incorporate on funding.

We would rather you hear this from us than work it out. It is also why every claim on this site carries a label. How to read this site

06Get in touch

Who we’d like to hear from.

  • Researchers

    Compare notes on compute-efficient agents, and on how to measure them honestly.

    The research

  • Grant and accelerator programmes, and investors

    We’ll walk you through our technical approach and costed build plan in confidence.

    Email the founder

  • Engineers

    Three roles are planned, in machine learning, systems, and simulation and robotics, subject to funding.

    The early team

  • Educators

    Small models that run offline on ordinary laptops, and education packages, are part of the plan.

    Start a conversation

  • Design partners

    Teams with long software projects, or data that must stay on their own infrastructure.

    Candid for partners

  • Everyone else

    Questions, corrections, or a result you think we got wrong. All of it is welcome.

    Contact